Create a proof of concept which shows how "2D" models can be trained from imaging volumes
Old description of data loader. Required for a 2D (slice based) UNet implementation. Pixel data pulled from Sonador is grouped by series and represents the entire volume. The data loader needs to be able to pull the members of the data collection and reliably create a new data structure composed of all slices.
Requirements:
- The loader needs to be able to reliably apply transformations to the volume (to provide a consistent orientation/projection, for example) and then apply slice based transformations.
- The loader should cache results to disk (or another appropriate storage) so that computationally intensive operations can be optimized. It should also play nicely with the cached dataset from MONAI.
Updated description (2023-0116): the PatchDataset
from MONAI can be used to sample 2D slice data from 3D volumes and conforms with the previous requirements. A sample notebook should be added to the examples repository which shows how to use PatchDataset
alongside Sonador AI. The MONAI tutorials repository contains a notebook which can probably be adapted.